AI and Ethics in Psychotherapy
This is post #2 in a series exploring AI in psychotherapy.
The ethical questions raised by bringing AI into therapy are not abstract, and they are not optional. They sit at the centre of whether this technology belongs anywhere near the consulting room, and they need to be faced directly rather than waved through on the strength of how useful the tools are. I want to set them out plainly here, because the profession is adopting these systems faster than it is thinking about them, and the gap between the two is where harm tends to live.
Three concerns matter most: data privacy, algorithmic bias, and transparency. Each of them touches the same thing in the end, which is trust, and trust is the ground that all therapeutic work stands on. Data privacy is the first duty People bring their most private material to therapy. They talk about abuse, about shame, about thoughts they have never said aloud to anyone, and they do so on the understanding that what they say stays in the room. AI tools complicate that understanding in ways many practitioners have not fully reckoned with. These systems run on data. They collect it, process it, and store it, and the sensitive nature of what gets shared in therapy makes that storage a genuine vulnerability rather than a technicality.
The honest position is that most people using these tools have no real idea where their words end up. Someone types the worst night of their life into a chatbot and has no sense of whether that text is held, for how long, by whom, or to what further use it might be put. GDPR and the AI Act 2024 set standards for how data must be handled, and they matter, but legislation written for data in general struggles to keep pace with the specific ways AI systems ingest and retain it. A breach in a therapeutic context carries a weight that an ordinary data leak does not. It can rupture the relationship the whole work depends on, and a ruptured therapeutic trust is not something you simply restore with an apology and a new password. So the duty here is clear. Any therapist using these tools has to know where the data goes, has to be able to explain it to a client in plain terms, and has to treat the answer as a condition of use rather than a detail to sort out later.
Bias is built in unless someone builds it out. A model is only ever as good as the data it learned from, and the data behind most of these systems skews heavily toward Western, white, middle-class populations. That skew becomes a clinical problem the moment the tool is used with anyone outside that default. In therapy, a person's culture and context shape everything about how they suffer and what will help them, and a system trained on a narrow slice of humanity will quietly mishandle the people it was never really built to understand.
This is one of the more insidious risks, because the tool gives no sign that it is doing it. It produces a confident recommendation, and the bias is folded invisibly into that confidence. A therapist who trusts the output without interrogating it can end up delivering skewed care to exactly the clients who are already least well served by the system, and call it evidence-based while doing so. Fairness in AI does not arrive on its own. It has to be deliberately engineered through diverse and representative training data, and where that work has not been done, the responsibility falls on the practitioner to assume the bias is present and to compensate for it.
Transparency is the price of trust The systems make decisions in ways neither the therapist nor the client can see into. This is the black box problem, and it bites harder in therapy than almost anywhere else. When a client asks why a particular recommendation was made, and the honest answer is that nobody can fully say, you have introduced something into the work that erodes the very thing therapy is built to provide. People will not engage with a process they cannot understand and cannot trust, and they are right not to.
Transparency is therefore an ethical requirement and a practical one at the same time. Clients need to understand how these systems reach the conclusions that affect their care, and they need to know what data is feeding those conclusions. The AI Act 2024 has begun to insist on this for high-risk systems in healthcare, demanding that decisions be explainable to people who are not engineers. That is the right direction. An AI tool that cannot explain itself has no place in sensitive clinical work, however impressive its outputs appear.
Accountability cannot be left vague Underneath all of this sits a question the field has not answered: when an AI system contributes to a bad outcome, who is responsible? If the tool steers a therapist toward a harmful recommendation, is the therapist liable for following it, or do the developers carry the fault for building it? The comfortable instinct is to leave this blurry, and the blurriness itself is dangerous, because responsibility that belongs to everyone tends to be carried by no one.
My own view is that the therapist remains accountable for clinical decisions, full stop. The tool is a tool, and using it does not transfer the duty of care to a piece of software. At the same time, developers carry a real and separate responsibility for the systems they release into clinical hands, and they should not be permitted to hide behind the therapist's final judgment as though their design choices were neutral. The profession needs clear protocols that name these lines before the difficult cases arrive, not after.
Where this leaves us The ethics here are not a hurdle to clear on the way to adoption. They are the substance of whether adoption is defensible at all. A therapist who uses these tools without confronting data, bias, transparency, and accountability is not being innovative. They are gambling with the trust their clients have placed in them, and that trust is the most precious thing they hold.
I am genuinely hopeful about what AI might come to offer this work. That hope only earns its place if it travels with a steady refusal to let the ethics slide. The tools will keep improving and the pressure to use them will keep growing, and the practitioners who do this well will be the ones who treat the ethical questions as the first thing to settle rather than the last.